(SSTA-NET)Speed and Short-Term Action Learning Networks for Indoor Wind Tunnel Skydiving Action Recognitions
Yuxiang Li, Liuyue Xu, Wei Zhang, Zhisheng Gao · 2024
Action recognition is widely used in video classification and motor behavior analysis. However, existing action recognition methods often overlook the importance of speed and distinct short-term actions. In particular, the similarity between various short-term actions during skydiving training poses a significant challenge for action recognition in indoor wind tunnel skydiving. We tackle these challenges by proposing a Speed and Short-Term Action Learning Network (SSTA-NET) for indoor wind tunnel skydiving action recognition. The network integrates a short-term action feature attention module that processes short-term action information within a single framework. To acquire dynamic motion data across varying movement types, SSTA-NET introduces a displacement map speed learning module that captures motion speed details in action videos. Our network also utilizes a video-efficient multi-scale attention module, designed to prioritize salient visual areas while reducing redundant spatio-temporal redundancy in the videos. SSTA-NET achieves an average improvement of 8.56% on the indoor wind tunnel skydiving dataset and 6.06% on the HMDB51 public dataset compared to other methods.